Politecnico di Torino

Porto Institutional Repository

[Article] Advances in quantitative muscle ultrasonography using texture

analysis of ultrasound images

Original Citation:
Molinari, Filippo; Caresio, Cristina; Acharya, U Rajendra; Mookiah, Muthu Rama Krishnan; Minetto,
Marco Alessandro (2015). Advances in quantitative muscle ultrasonography using texture analysis
of ultrasound images. In: ULTRASOUND IN MEDICINE AND BIOLOGY, vol. 41 n. 9, pp. 2520-
2532. - ISSN 0301-5629
Availability:
This version is available at : http://porto.polito.it/2621825/ since: July 2016
Publisher:
Elsevier
Published version:
DOI:10.1016/j.ultrasmedbio.2015.04.021
Terms of use:
This article is made available under terms and conditions applicable to Open Access Policy Article
("Public - All rights reserved") , as described at http://porto.polito.it/terms_and_conditions.
html
Porto, the institutional repository of the Politecnico di Torino, is provided by the University Library
and the IT-Services. The aim is to enable open access to all the world. Please share with us how
this access benefits you. Your story matters.

(Article begins on next page)

 Ultrasound in Med. & Biol., Vol. 41, No. 9, pp. 25202532, 2015
Copyright
2015 World Federation for Ultrasound in Medicine & Biology
Printed in the USA. All rights reserved
0301-5629/$ - see front matter

http://dx.doi.org/10.1016/j.ultrasmedbio.2015.04.021

d Original Contribution

ADVANCES IN QUANTITATIVE MUSCLE ULTRASONOGRAPHY USING TEXTURE

ANALYSIS OF ULTRASOUND IMAGES

FILIPPO MOLINARI,* CRISTINA CARESIO,*y U. RAJENDRA ACHARYA,zx

MUTHU RAMA KRISHNAN MOOKIAH,z and MARCO ALESSANDRO MINETTOyk
* Biolab, Department of Electronics and Telecommunications, Politecnico di Torino, Turin, Italy; y Division of Endocrinology,
Diabetology and Metabolism, Department of Medical Sciences, University of Turin, Turin, Italy; z Department of Electronics
and Computer Engineering, Ngee Ann Polytechnic, Singapore; x Department of Biomedical Engineering, SIM University,
Singapore; and k Division of Physical Medicine and Rehabilitation, Department of Surgical Sciences, University of Turin, Turin,
Italy
(Received 22 August 2014; revised 21 March 2015; in final form 27 April 2015)

AbstractMusculoskeletal ultrasound imaging can be used to investigate the skeletal muscle structure in terms of
architecture (thickness, cross-sectional area, fascicle length and fascicle pennation angle) and texture. Gray-scale
analysis is commonly used to characterize transverse scans of the muscle. Gray mean value is used to distinguish
between normal and pathologic muscles, but it depends on the image acquisition system and its settings. In this
study, quantitative ultrasonography was performed on five muscles (biceps brachii, vastus lateralis, rectus femoris,
medial gastrocnemius and tibialis anterior) of 20 healthy patients (10 women, 10 men) to assess the characteriza-
tion performance of higher-order texture descriptors to differentiate genders and muscle types. A total of 53 fea-
tures (7 first-order descriptors, 24 Haralick features, 20 Galloway features and 2 local binary pattern features)
were extracted from each muscle region of interest (ROI) and were used to perform the multivariate linear regres-
sion analysis (MANOVA). Our results show that first-order descriptors, Haralick features (energy, entropy and
correlation measured along different angles) and local binary pattern (LBP) energy and entropy were highly
linked to the gender, whereas Haralick entropy and symmetry, Galloway texture descriptors and LBP entropy
helped to distinguish muscle types. Hence, the combination of first-order and higher-order texture descriptors
(Haralick, Galloway and LBP) can be used to discriminate gender and muscle types. Therefore, multi-texture anal-
ysis may be useful to investigate muscle damage and myopathic disorders. (E-mail: filippo.molinari@polito.
it)
2015 World Federation for Ultrasound in Medicine & Biology.
Key Words: Muscle ultrasonography, Texture analysis, Muscle characterization, Musculoskeletal ultrasound, Har-
alick features, Galloway features, Local binary pattern, Gray-scale mean value, MANOVA.

Ultrasound imaging is proven to be effective in the inves-
tigation of the skeletal muscle structure (Pillen et al.
2008; Pillen and van Alfen 2011; Walker et al. 2004).
The main advantages of ultrasounds are portability, low
associated costs of the examination and non-
invasiveness of the method. Moreover, the acoustic power
levels used in diagnostic equipment minimize the proba-
bility of biological negative effects. Ultrasound imaging
is, however, an operator-dependent technique. In order
Address correspondence to: Prof. Filippo Molinari, Biolab,
Department of Electronics and Telecommunications, Politecnico di
Torino, Corso Duca degli Abruzzi, 24 - 10129 Torino, Italy. E-mail:
filippo.molinari@polito.it
to lower the intra- and inter-reader variability, a quantita-
tive approach is needed.
The quantitative features most commonly extracted
from ultrasound images to investigate muscle size are
cross-sectional area, thickness, fascicle length and
fascicle pennation angle (Chow et al. 2000; Narici et al.
1996; Reeves et al. 2004). The muscle quality is
commonly assessed through the quantification of the
mean echo intensity by gray-scale analysis of a region
of interest (ROI). This numerical parameter is highly
dependent on the ultrasound scanner settings (Pillen
et al. 2009b; Zaidman et al. 2008).
Contrary to the mean echo intensity and other first-
order descriptors, higher-order texture features that can
be extracted from ultrasound images are intensity
invariant (Acharya et al. 2012d, 2012e, 2012f) and have

2520
 Muscle ultrasonography texture analysis d F. MOLINARI et al.
already proven informative in the analysis of
intramuscular fat content in animals (Kim et al. 1998),
as well as in the characterization of arterial surface rough-
ness (Acharya et al. 2012a; Niu et al. 2013), breast (Singh
and Singh 2010) and ovarian tumors (Acharya et al.
2012c, 2013), thyroid lesions (Acharya et al. 2012b,
2014a) and liver images (Acharya et al. 2015; Gao
et al. 2014) in human studies. Moreover, in a recent
review focused on thyroid cancer, Acharya et al. 2014
reported that high characterization performance can be
achieved only when sonographic features from the
ultrasound images are merged to non-clinical features
extracted from the ultrasound images using statistical
and data mining techniques (Acharya et al. 2014b).
They have also found that higher-order and non-linear de-
scriptors offer better characterization performance than
histogram-based parameters (Acharya et al. 2012c,
2014b).
To the best of our knowledge, in previous studies,
only linear and first-order descriptors are used to charac-
terize the texture of different skeletal muscles. In this pa-
per, we characterized the image texture of five skeletal
muscles of healthy men and women using different
texture features. We show that Haralick features (sec-
ond-order statistical descriptors), Galloway features and
texture descriptors based on the local binary pattern
(LBP) are unique for gender and muscle type. It can be
seen from our results that the texture features are superior
to the first order descriptors (based on the echo intensity
histogram) in classifying the muscle type.
METHODS
Patients
Twenty healthy volunteers (10 females: aged 26.0 6
2.3 y; body mass index 20.7 6 2.2 kg/m2 and 10 males,
aged 30.2 6 5.6 y; body mass index 23.3 6 2.6 kg/m2)
participated in this study. Health status was assessed by
medical history, clinical examination and electrocardio-
gram. The Waterloo Handedness and Footedness Ques-
tionnairesRevised (Elias et al. 1998) was used to
assess side dominance. Among them, three patients
were left-side dominant. Before participating in the study,
the patients were instructed about the aims and then they
signed a written informed consent. The study conformed
to the guidelines of the Declaration of Helsinki and was
approved by the local ethical committee.
Ultrasound procedures and equipment
During a single experimental session, we acquired
ultrasound B-mode images of the following five muscles
from each subject: biceps brachii, vastus lateralis, rectus
femoris, medial gastrocnemius and tibialis anterior. Im-
ages were acquired on both sides of the patients.
2521
The same experienced sonographer (MAM) con-
ducted the clinical examinations and acquired all the im-
ages. Three consecutive scans were acquired in the
transverse plane of each muscle. After each scan, the sub-
ject moved and then the transducer was repositioned. To
increase the repeatability of the acquisitions and to ensure
that the insonation was orthogonal to the bone, the
optimal insonation angle was selected by maximizing
the representation of the bone boundary.
The medial gastrocnemius was insonated with the
patients in prone position, whereas for all other muscles
the supine position was maintained. In all measurements,
the arms and legs were extended and the patients were
asked to completely relax their muscles. Ultrasound
coupling gel (Ultrasound transmission gel, REF:
907137475, PBpharma, Torino, Italy) was used to ensure
optimal image quality while limiting the transducer pres-
sure on the skin. All scans were performed by placing the
transducer in correspondence to the largest muscle diam-
eter at the following anatomic sites: The biceps brachii
was measured at two thirds of the distance from the acro-
mion to the antecubital crease; the rectus femoris halfway
along the line from the anterosuperior iliac spine to the
superior border of the patella; the vastus lateralis halfway
along the line from the anterosuperior iliac spine to the
superolateral border of the patella; the tibialis anterior
at one quarter of the distance from the inferior border
of the patella to the lateral malleolus; the medial gastroc-
nemius from the mid-sagittal line of the muscle, midway
between the proximal and distal tendon insertions.
We used a MyLab Twice ultrasound device (Esaote,
Genoa, Italy) equipped by a linear-array transducer (code
LA533) with a bandwidth from 313 MHz. Gain was set
at 50% of the range, dynamic image compression was
turned off, and time gain compensation was maintained
in the same (neutral) position for all depths. All system-
setting parameters were kept constant throughout the
study for each subject. The depth setting (initially set at
44 mm) was adapted for each participant during examina-
tion in order to display the entire muscle. The conversion
factor was equal to 0.92 mm/pixel. The pictures in
DICOM format were transferred to a computer for offline
processing.
Texture feature extraction
All images were visually inspected and analyzed by
the same experienced operator (CC), who positioned a
ROI in each image as shown in Figure 1. One ROI was
chosen in the median portion of the biceps brachii, vastus
lateralis and medial gastrocnemius, whereas two equal-
sized ROIs were chosen in the rectus femoris (Fig. 1a)
and tibialis anterior (Fig. 1c) to include most of the muscle
without the central aponeurosis (white arrow in Fig. 1a) and
the internal fascia (white arrow in Fig. 1c). Figure 1b
2522 Ultrasound in Medicine and Biology Volume 41, Number 9, 2015

Fig. 1. ROI positioning in the transversal B-mode images. (a) rectus femoris; (b) vastus lateralis; (c) tibialis anterior; (d)
medial gastrocnemius. The white arrows in panels (a) and (c) indicate, respectively, the central aponeurosis and the in-
ternal fascia that were never included in the ROI. ROI 5 region of interest.

depicts the ROI positioning for the vastus lateralis and Haralick features. The Haralick features (also
Figure 1d for the medial gastrocnemius. The two ROIs of called second-order statistical descriptors) are based
the rectus femoris are indicated as medialis and lateralis on the gray level co-occurrence matrix (GLCM)
and abbreviated as RFMed and RFLat, respectively. The ti- (Haralick et al. 1973). The GLCM is a square matrix
bialis anterior ROIs are indicated as superior and inferior with dimension equal to the number of gray levels in
and abbreviated as TASup and TAInf, respectively. The the image. Let C be the matrix containing the GLCM.
dimension and the position of the ROIs were chosen to The element C(i,j) measures the number of times in
be the same for each muscle of all patients in order to which a pixel of given gray level i is found adjacent
make the extracted features independent of ROI size. The to a pixel of gray level j. Because two pixels can be
following ROI areas were considered for each muscle: adjacent in vertical, horizontal and in the two diagonal
286 mm2 for the biceps brachii, 144 mm2 for the rectus directions, we computed C(i,j) for the four angles 0 ,
femoris, 338 mm2 for the vastus lateralis, 68 mm2 for the 45 , 90 and 135 . Figure 2 shows an example of
tibialis anterior, 338 mm2 for the medial gastrocnemius.
Different features extracted from the ROIs. The
Table 1. Mathematical description of first-order
mean of six measurements (one measurement per ROI statistical features. Input muscle region of interest (ROI)
for both sides) was used for comparison among muscles is denoted by I(x,y)
and between genders. All the texture parameters were
Feature Description
computed by custom developed software in MATLAB
PM PN
(The MathWorks, Natick, MA, USA). Integrated optical density (IOD) IOD 5 x 5 1 y 5 1 Ix; y
P PN Ix;y
Mean (m) m5 M x51 y 5 1 M3N
First-order statistical descriptors. Based on the first r
P P
Standard deviation (s) M N
fIx;y2mg2
order statistics, the following seven features were ex- s5 x51 y51

tracted from the image ROIs: integrated optical density, PM PN M3N

Variance (s2) fIx;y2mg2
s2 5 x51 y51
mean, standard deviation, variance, skewness, kurtosis PM P
M3N
N
Skewness (Sk) fIx;y2mg3
and energy. Table 1 presents the mathematical description Sk 5 M3N
1 x51 y51

PM P N
s3
of these features. The energy feature is denoted as En- Kurtosis (Kt) fIx;y2mg4
Kt 5 M3N
1 x51 y51
s
P PN
4
ergy1 to avoid confusion with the same second-order Energy1 (E1) E1 5 M 2
x51 y 5 1 Ix; y
parameter (described in the next section).
 Muscle ultrasonography texture analysis d F. MOLINARI et al. 2523

Fig. 2. Schematic representation of the computation of the GLCM. (a) Numerical values corresponding to the pixel in-
tensities. (b) Associated GLCM. (c) Gray scale representation of the image in (a). (d) Gray scale representation of the
GLCM. The red circles in (a) depict the adjacencies (1,6), which correspond to the number circled in red in panel (b).
GLCM 5 gray level co-occurrence matrix.

construction of the GLCM for the horizontal direction: (Galloway 1975) observed that in coarse texture, long
Figure 2a depicts the numerical values representing gray level runs may exist more frequently compared
the pixels of an image that is rendered in a linear with fine texture, which generally contains short runs.
gray-scale in Figure 2c. The corresponding GLCM is This means that we expect longer sequences of pixels
numerically computed in Figure 2b and the gray-scale with same intensity in coarse images but only short
representation of the GLCM is shown in Figure 2d. runs in fine textures. The Galloway features are then
The red circles in Figure 2a indicate the horizontal mathematical descriptors of the runs of the RLM: short
adjacencies of the pixel i 5 1 and of the pixel j 5 6. run emphasis, long run emphasis, gray-level non-unifor-
There are four occurrences of this adjacency, hence mity, run length non-uniformity and run percentage.
the corresponding pixel C(1,6) in Figure 2b has a value Table 3 reports the mathematical description of the fea-
equal to 4. tures that were calculated from the RLM. Because each
The Haralick features are the mathematical descrip-
tors of the GLCM (Fig. 2d). We computed the following
six features: symmetry, contrast, homogeneity, entropy, Table 2. Mathematical description of the Haralick
energy and correlation. Because each feature is computed features
along 4 angular directions, we have 24 descriptors per Haralick feature Description
ROI. These second-order features are mathematically P PN21
defined in Table 2. The full mathematical details about Symmetry (Isym) Isym 5 12 N21 ji2jjPi; j
PN21i 520 Pj N5 0 PN
the Haralick features are reported in the Appendix. Contrast (Icon) Icon 5 n 5 0 n f i 5 0 j 5 0 Pi; jg
P PN21 1
Homogeneity (Ihmg) Ihmg 5 N21i50 j 5 0 11i2j2 Pi; j
Galloway features. The Galloway features PN21 PN21
Entropy (IEntr) IEntr 5 2 i 5 0 j 5 0 Pi; jlogPi; j
(Galloway 1975) are based on the run length matrix P PN21
Energy (IEnrg) IEnrg 5 N21 j 5 0 Pi; j
2
(RLM) R. In a RLM, the pixel R(i,j) contains the number PN21i 5P
0
N21
Correlation* (Icor) i;jPi;j2m m
Icor 5 i 5 0 j 5s0x sy
x y
of pixels with run length j and intensity i in a given direc-
tion. The RLM has a number of rows equal to the number
* sx ; sy ; mx ; my are the standard deviations and means of Px ; Py , which
of gray levels in the image and a number of columns equal are the partial probability density functions. px i 5 ith entry in the mar-
to the maximum length of the run length. Galloway ginalprobability matrix obtained by summing the rows of Pi; j.
2524 Ultrasound in Medicine and Biology Volume 41, Number 9, 2015

Table 3. Mathematical description of the Galloway (Fig. 3d) is used to compute the LPB, which is the sum of
features all the pixels (in the case of Fig. 3d, the LBP 5 216).
Galloway feature Description Hence, each pixel is assigned a LBP value. Acharya
PNg PNr Ri;j et al. reported that two powerful descriptors of the LBP
Short run emphasis (SRE) image are the energy and entropy of the LBP distribution
SRE 5 PNig5 1PNjr5 1 j
2

Ri;j
PN g PN r 2
i51 j51 (Acharya et al. 2012d). Particularly, the indicators based
Long run emphasis (LRE) j Ri;j
LRE 5 Pi 5Ng1 Pj 5Nr1 on LBP have lower values for more homogeneous images
Ri;j
P i51
Ng PNr
j51
and assume higher values for more inhomogeneous im-
Gray-level non- Ri;j2
uniformity (GLNU) GLNU 5 Pi 5N1g PjN5r 1 ages. The full mathematical details of the LBP computa-
Ri;j
PNri 5 1PNjg5 1 tion method we adopted are reported in the Appendix.
Run length non- i 5 1 Ri;j2
uniformity (RLNU) RLNU 5 PNg PNr
j51

Ri;j
PNg Pi 5Nr1 j 5 1
Run percentage (RP) Ri;j Statistical analysis and classification
RP 5 i 5 1Ng $Mj 5r 1
The Shapiro-Wilk test was used to assess the
Ng represents the number of gray values in the image (i.e., the number normality of the variable distributions. Normally distrib-
of rows of the matrix R). Nr represents the number of runs (i.e., the num- uted data were analyzed using the Students t-test,
ber of columns of the R matrix). whereas non-normally distributed data were analyzed us-
ing the Mann-Whitney U test. The Kruskal-Wallis AN-
feature was computed along four different angles (0 , 45 , OVA (followed by Dunns post hoc test) was adopted
90 and 135 ), we have extracted 20 Galloway features for comparing the features value among the different
per ROI. muscles.
The overall number of texture descriptors was equal
Local Binary Pattern Features. The LBP was intro- to 53 per ROI (7 first-order descriptors, 24 Haralick fea-
duced by Ojala et al. in the field of texture analysis and tures, 20 Galloway features and 2 LBP features). Because
face recognition (Ojala et al. 1996, 2002). Basically, by we had 5 muscles and 7 ROIs (the RF and TA were
means of this technique, a LBP value is assigned to a measured in two different ROIs), and because we
pixel neighborhood on the basis of the comparison of measured the dominant and non-dominant side, each sub-
the pixels intensities with the intensity of the central ject was represented by 742 features. Multivariate anal-
pixel. With reference to Figure 3, lets consider the cen- ysis of variance (MANOVA) was used to test the
tral pixel of the 3 3 3 neighborhood in Figure 3a. The in- equality of the means among groups. We tested the
tensity of the central pixel is used to threshold the texture feature values against gender and against the mus-
neighborhood and a pixel is assigned the value 0 if its in- cle type. To avoid singularities in the observation matrix,
tensity is lower than the center or 1 otherwise (Fig. 3b). before performing the MANOVA analysis, collinear vari-
Then the binarized image is multiplied by the image con- ables were removed by computing the Wilks lambda,
taining the powers of 2 (Fig. 3c) and the resulting image which is defined as the ratio between the determinant of
the within group variance matrix and the sum between
the determinants of the within and between matrices.
a b Conceptually, the Wilks lambda is the proportion of
the total variance in the discriminant scores not explained
3 4 1 0 0 0 by differences among the groups. The lambda ranges
7 5 9 1 1 from 0 to 1 and the lower is the value, the more discrim-
inant between groups is the associated variable (Costanza
8 2 7 1 0 1 and Afifi 1979). The optimal lambda value for our data set
was found to be 0.35. Higher values lead to insufficient
c d removal of collinear variables, whereas lower values dis-
carded an excessive number of variables. The dimension
1 2 4 0 0 0 of the MANOVA was used to assess the number of groups
the data belong to. For example, a dimension equal to
128 8 128 8
zero indicates that it is not possible to reject the hypoth-
64 32 16 64 0 16 esis that all the patients belong to the same group,
whereas a dimension equal to 1 indicates that the patients
Fig. 3. Schematic representation of the LBP process. (a) Orig- can be divided into two groups. Classification of the pa-
inal pixel intensities. (b) Binarization by thresholding. (c)
Power-of-two weights. (d) Final LBP neighborhood. The overall tients on the basis of either the gender or muscle type
LBP value assigned to the central pixel is the sum of the was obtained by linear regression analysis performed
numbers in (d), which equals 216. LBP 5 local binary pattern. on the most significant features, as revealed by
 Muscle ultrasonography texture analysis d F. MOLINARI et al.
MANOVA. All the continuous data variables were repre-
sented by means 6 standard deviation (SD). The statisti-
cal significance is set to p 5 0.05. Statistical analyses
were performed by R (www.r-project.org) and MATLAB.
RESULTS
Comparisons between sides and genders
All the texture features of the dominant side were
correlated to the non-dominant side (p , 0.001). The
lowest value of the Wilks lambda was 0.875. Accord-
ingly, the MANOVA analysis was not significant when
the side dominance was the dependent variable. There-
fore, we averaged the variables of the two sides: The total
number of variables was equal to 371 for each of the
patients.
When the gender was considered as the dependent
variable, after removing the collinear variables, 38 fea-
tures were left. The MANOVA dimension of the group
means was equal to 1 (p , 0.001). The dimensionality
of the MANOVA was important to understand how sam-
ples were distributed on the hyperplane of the canonical
variables. The canonical variables are linear combina-
tions of the original features and are built in order to
maximize the variance among groups. Because the ca-
nonical variables are ordered with decreasing explained
variance, the dimension of 1 ensured that only one canon-
ical variable (the first) is enough to separate the patients
on the basis of gender. We plotted the first and second ca-
nonical variable for each subject (Fig. 4), and we indi-
cated the males by a full symbol (circle) and females
by an empty one and kept the same notation throughout
the paper. The graph indicates that the first canonical var-
iable was discriminant for gender. The left column of
Table 4 reports the 10 variables (i.e., the image features)
with highest weight on the first canonical variable (i.e.,
the most discriminant between the two genders). First-
order and Galloway features were not significant (p .
0.2). Haralick features (energy, entropy and correlation),
LBP energy and entropy were significantly higher (Hara-
lick energy, p , 0.01) and lower (all other features, p ,
0.001) in men compared with women (Fig. 5). By using
those 10 most discriminant features, we performed a clas-
sification of the patients based on the linear regression.
All the patients were correctly classified, with sensitivity
and specificity of 100% (correlation coefficient equal to 1
and 95% CI 0.981) and an area under the receiver oper-
ating curve (AUROC) equal to 1.
Comparison among muscles
When the muscle was considered as the dependent
variable, after removing the collinear variables, 43 fea-
tures were left. The MANOVA dimension was 6 (p ,
0.02), thus we could consider that samples belonged
2525
Fig. 4. Representation of the patients in the plane of the first two
canonical variables obtained by MANOVA. The features al-
lowed for a clean-cut separation of the patients on the basis of
the gender. The full symbols represent the males (data from
two male patients are overlapped), the empty symbols the
females.
to 7 different groups. Figure 6a shows the different mus-
cles plotted in the plane of the first two canonical vari-
ables, whereas Figure 6b shows the same distribution in
the plane of the first and third canonical variables. The
right column of Table 4 reports the 10 original features
that were most discriminant among the different mus-
cles. The most important features were the Galloway
ones (gray-level non-uniformity, run length non-
uniformity, run percentage, short run emphasis), the
Haralick ones (entropy and symmetry) and the
LBPentropy. No first-order features are listed. Figure 7
shows the gray-level non-uniformity (GLNU, Fig. 7a),
the run length non-uniformity (RLNU, Fig. 7b), the
LBPentropy (Fig. 7c) and the Haralick entropy (Fig. 7d)
for the different muscles and ROIs. No significant
Table 4. Image features that are the most discriminant
between the two genders and among the five muscles in
the MANOVA analysis
Most discriminant features Most discriminant features
for gender for muscle type
Haralick energy GLNU (q 5 0 , 45 , 90 )
(q 5 0 , 45 , 90 , 135 )
Haralick entropy (q 5 90 , 135 ) Haralick entropy (q 5 135 )
LBPenergy LBPentropy
Haralick correlation (q 5 45 , 90 ) RLNU (q 5 135 )
LBPentropy Haralick symmetry (q 5 90 )
RP (q 5 90 )
SRE (q 5 0 , 90 )
LBPenergy2LBPentropy 5 local binary pattern features; GLNU 5
gray-level non-uniformity; RLNU 5 run length non-uniformity; RP 5
run percentage; SRE 5 short run emphasis.
The features are listed in order of decreasing weight in the MANOVA
canonical variables.
2526 Ultrasound in Medicine and Biology Volume 41, Number 9, 2015

Fig. 5. Distribution of the four most discriminant features for gender in the sample groups. The full symbols represent the
males, the empty symbols the females. The red circles indicate the biceps brachii muscle, the down triangles the inferior
ROI of the tibialis anterior. In each graph, the mean value and the interval corresponding to 6 2 standard deviations is
represented. (a) Haralick energy (along the angular direction q 5 135 ). (b) Haralick entropy (q 5 135 ). (c) Haralick
correlation (q 5 90 ). (d) Local binary pattern (LBP) entropy.

difference in the texture features was observed between
the different ROIs selected in the rectus femoris and ti-
bialis anterior muscles (p . 0.05 for all comparisons).
Galloway features (GLNU and RLNU) and the Haralick
entropy resulted higher (p , 0.05 for all comparisons)
in biceps brachii, medial gastrocnemius and vastus later-
alis muscles compared with rectus femoris and tibialis
anterior muscles. It can be noticed that the texture de-
scriptors considered alone do not have a very high
discriminatory power.
The performance in classifying different muscles us-
ing the linear regression is reported in Table 5 (leftmost
half). The average sensitivity was equal to 76.4 6
21.9%, the specificity to 97.7 6 1.9%, the correlation co-
efficient was 0.98 (95% confidence interval 0.931), and
the AUROC to 0.976 6 0.026. The two ROIs of the tibia-
lis anterior were perfectly classified (AUROC 5 1), and
the worst performance was obtained in classifying the
rectus femoris (AUROC 5 0.936).
Comparisons between genders and among muscles
based on the first-order descriptors
To further compare the performance of the Hara-
lick, Galloway and LBP features with the first-order
descriptors, we performed the MANOVA analysis and
the classification using only the 7 first-order features.
Figure 8a represents the patients on the basis of the
gender as returned by the first two canonical variables
of the MANOVA analysis. The dimension obtained by
the MANOVA was 1 (p , 0.02), thus allowing for the
separation of the samples in two groups. A distinction
of the two groups is still evident, even though not
defined as in Figure 4 for Haralick and LBP features.
Accordingly, the first-order features were significantly
 Muscle ultrasonography texture analysis d F. MOLINARI et al.
Fig. 6. Representation of the patients (a) in the plane of the first
two canonical variables, and (b) in the plane of the first and third
canonical variable obtained by MANOVA. Different symbols
and colors are used for the different muscles. The full symbols
represent the males, the empty symbols the females. BB 5 bi-
ceps brachii; MG 5 medial gastrocnemius; VL 5 vastus later-
alis; RF 5 rectus femoris; TA 5 tibialis anterior.
different between men and women. It can be seen from
Figure 9 that the mean echo intensity of four muscles
was higher for women than men. When we classified
the patients on the basis of the gender using only the
seven first-order features, we obtained a 100% sensi-
tivity and specificity and an AUROC equal to 1. Hence,
the use of Haralick and LBP features did not improve
the description of patients on the basis of gender.
Similarly, we classified the muscles using only the
seven first-order descriptors and we did not get clear sep-
aration among the muscles (Fig. 8b). As the muscles are
not separated as in Figure 6, we expected a lower classi-
fication performance when first-order descriptors were
used to differentiate among the five muscles. The dimen-
2527
sion obtained by the MANOVA was 6 (p , 0.05). The
rightmost part of Table 5 reports the classification perfor-
mance we obtained when the linear regression was per-
formed using the seven first-order features. The average
sensitivity of 46.4 6 32.3%, the specificity of 97.3 6
1.92% and AUROC of 0.907 6 0.081 were obtained.
The sensitivity and the AUROC were significantly (p ,
0.01) lower than Haralick, Galloway and LBP features.
Consistently, the first-order features were statistically
comparable among the five muscles (p . 0.05). In other
words, the overall performance was lower than when
higher-order descriptors were used, but this decreased
performance was consistent for all the muscles.
Briefly, it was possible to differentiate between the
two genders on the basis of the first-order descriptors,
whereas the Haralick, Galloway and LBP features classi-
fied the images of different muscles with a better perfor-
mance than the first-order features.
DISCUSSION
In the present study, quantitative ultrasonography
was performed in five muscles (biceps brachii, rectus
femoris, vastus lateralis, tibialis anterior, medial gastroc-
nemius) of 20 healthy patients to assess the characteriza-
tion performance of higher-order texture descriptors in
the differentiation between genders and among muscles.
Our results show that first-order descriptors, Haralick fea-
tures (energy, entropy and correlation measured along
different angles), LBP energy and entropy were highly
linked to the gender, whereas Haralick entropy and sym-
metry, Galloway texture descriptors and LBP entropy
helped to distinguish different types of muscles. To the
best of our knowledge, this is the first study that used
higher-order textural descriptors to characterize the hu-
man muscles using ultrasound images.
Previous studies have found that the first-order de-
scriptors (i.e., mean or median echo intensity) help to
study the echogenicity of different genders and skeletal
muscles (Arts et al. 2010) and muscle adaptations to
physical training (Radaelli et al. 2012). Moreover, the
quantification of the muscle echo intensity can charac-
terize the disruption of the tissue architecture that occurs
in myopathic, neuropathic, and damaged muscles. Neuro-
muscular disorders and myopathic conditions are associ-
ated to fat and collagen infiltration (Arts et al. 2010;
Pillen et al. 2008) and/or interstitial edema (Fujikake
et al. 2009). These histologic changes increase the reflec-
tion of the ultrasound beam and result in increased echo
intensity (Pillen et al. 2008; Walker et al. 2004).
Despite these results, the muscle echo intensity pre-
sents a major criticism that limits its clinical use: it is
dependent on the settings of the ultrasound scanner and
is therefore different for each ultrasound device used.
2528 Ultrasound in Medicine and Biology Volume 41, Number 9, 2015

Fig. 7. Distribution of four discriminant features for muscle type in the sample groups. The full symbols represent the
males, the empty symbols the females. In each graph, the mean value and the interval corresponding to 6 two standard
deviations is represented. (a) GLNU (along the angular direction q 5 90 ). (b) Run length non-uniformity (RLNU,
q 5 135 ). (c) LBP entropy. (d) Haralick entropy (q 5 135 ). GLNU 5 gray-level non-uniformity; LBP 5 local
binary pattern.

To improve the reproducibility of the echo intensity,
Zaidman et al. studied a calibration procedure
(Zaidman et al. 2008, 2012) and proposed a calibrated
muscle backscatter index, which was adopted in the
analysis of muscular dystrophy (Zaidman et al. 2010).
The need for a calibration procedure is, however, imprac-
tical in the clinical environment. Also, as reported by the
authors, the quality of the calibration depends on the ROI
size and location inside the muscle (Zaidman et al. 2012).
Another approach for the comparison of echo intensity
values between different ultrasound devices was pro-
posed by Pillen et al. (2009b), who adopted a conversion
equation (based on standardized measurements of the
same samples made with two ultrasound devices) to
transpose the echo intensity values obtained with one de-
vice to another device.
Another criticism of the muscle echo intensity is that
it is able to capture only the overall change in the bright-
ness of the muscle, but it cannot measure the actual
texture. Few studies have attempted to develop quantifi-
cation methods that can overcome the limitation of
echo intensity in measuring the image texture. Maurits
et al. proposed quantitative variables obtained by density
analysis (such as muscle inhomogeneity and white-area
index, which measures the presence of patches of high
echogenicity) to differentiate between myopathies and
neuropathies (Maurits et al. 2003) and to distinguish be-
tween healthy muscles and neuromuscular diseases
(Maurits et al. 2004). The rationale was that highly inho-
mogeneous muscles showed a higher number of bright
pixels in the gradient compared with homogeneous mus-
cles. The white-area index is anyway correlated to the
overall brightness of the ultrasound image; thus, it suffers
from the same limitations of the mean echo intensity.
Gdynia et al. used the muscle echo intensity, the first-
order entropy, and the fractal dimension to analyze the
ultrasound images of tibialis anterior and medial gastroc-
nemius muscles of healthy patients and patients affected
 Muscle ultrasonography texture analysis d F. MOLINARI et al. 2529

Table 5. Classification performance based on the 10 most significant features as obtained by the MANOVA analysis (leftmost
half of the table) and based on the first-order features (rightmost half of the table)
Ten most significant features First order features only

Sensitivity (%) Specificity (%) AUROC Sensitivity (%) Specificity (%) AUROC

BB 80.0 97.5 0.984 100 100 1

MG 85.0 98.3 0.990 50.0 95.0 0.932
VL 45.0 95.0 0.942 10.0 95.8 0.851
RFMed 75.0 97.5 0.980 65.0 98.3 0.930
RFLat 50.0 95.8 0.936 45.0 95.8 0.910
TAInf 100 100 1 50.0 96.7 0.958
TASup 100 100 1 5.0 99.2 0.765

AUROC 5 area under the receiver operating curve; BB 5 biceps brachii; MG 5 medial gastrocnemius; VL 5 vastus lateralis; RFMed2RFLat 5 me-
diolateral portion of the rectus femoris; TAInf2TASup 5 inferosuperior portion of the tibialis anterior.

by myopathies and motor neuron disorders (Gdynia et al.
2009). They found that the parameters were able to distin-
guish between healthy and pathologic muscles but not be-
tween muscles affected by different pathologic
conditions. Entropy filtering, combined with vibration
elastography, was used in another study on the myofascial
trigger points of the upper trapezius muscle (Turo et al.
2013). The combination of the two measurements lead
to overall good performance (sensitivity of about 70%

Fig. 8. Results of the MANOVA analysis when only first order

features are used to distinguish between genders and muscle
types. (a) Representation of the patients in function of the
gender. (b) Representation of the patients in function of the mus-
cle type. BB 5 biceps brachii; MG 5 medial gastrocnemius;
VL 5 vastus lateralis; RF 5 rectus femoris; TA 5 tibialis
anterior.
Fig. 9. Differences in mean echo intensity (first-order feature)
of four muscles between the two genders. The full symbols
represent the males, the empty symbols the females. In each
graph, the mean value and the interval corresponding to 62
standard deviations is represented.
2530 Ultrasound in Medicine and Biology
and specificity of 80%) in discriminating myofascial
trigger points from normal tissue. Both the studies found
that entropy analysis can extract more information from
the B-mode images than the mean echo intensity. Howev-
er, the measurement of the first-order entropy is based on
the intensity histogram distribution of the pixels. Thus,
the texture information present in the image and reflect-
ing the muscle microstructure is not fully captured.
On the contrary, the image texture analysis enables
to overcome both limitations of the first-order descriptors
(Acharya et al. 2012a, 2012b, 2012c). In the present
study, we used both first-order descriptors and higher-
order texture descriptors to discriminate between genders
and different skeletal muscles of healthy patients.
The main findings of the present study are the
following:
1. First-order and texture descriptors are comparable to
the dominant and non-dominant side of each muscle.
2. First-order features help to distinguish the two
genders.
3. Galloway features quantify the coarseness of an image
along a given direction (i.e., describe the spatial
changes in the local acoustic impedance of the tissues)
and enable to distinguish different types of muscles.
4. Haralick and LBP features (in particular, Haralick en-
tropy and LBP entropy) quantify the overall (Haralick
entropy) and local (LBP entropy) image homogeneity
and distinguish both gender and muscle types.
A previous study on the echo-intensity of the tibialis
anterior and of the upper limb muscles found that the first-
order descriptors had different values in the dominant and
non-dominant side (Arts et al. 2010). In our study, the
first-order descriptors failed to differentiate between the
dominant and non-dominant. Hence, further studies are
required to examine the dependence of muscle echo in-
tensity (and texture as well) on the side dominance.
Our observation of women indicating higher echo
intensity compared with men is in agreement with previ-
ous findings (Arts et al. 2010; Caresio et al. 2014). But the
observation that the image homogeneity (as quantified by
Haralick and LBP features) is lower in women compared
with men and directional coarseness (as quantified by
Galloway features) difference among muscles is the
original contribution of the present study. Although no
muscle biopsies were performed in this study, the
observed gender variability in first-order descriptors,
Haralick features and LBP features may be related to
the higher degree of fibrous and adipose tissue present
in muscles of women compared with men (Caresio
et al. 2014; Pillen et al. 2009a). The inter-muscle vari-
ability in directional coarseness probably resulted from
different proportions of fibrous tissue and/or from
different architectural features of the fascicles (i.e., their
Volume 41, Number 9, 2015
lengths, their arrangement relative to one another and
their alignment relative to the force-generating axis).
Given these possible relations between textural features
and adipose and fibrous tissue content and distribution,
future studies are required to investigate the type of
textural features that enable to detect differences between
healthy and pathologic muscles. For example, conditions
associated to interstitial edema in a first phase and
collagen infiltration in a subsequent phase (e.g., muscle
injuries) can be longitudinally assessed by quantitative
muscle ultrasonography for monitoring and prognosis
of the myopathic process. We found in a series of endo-
crine patients affected by steroid myopathy that the mus-
cle texture impairment occurs earlier than the muscle size
reduction (unpublished observations). Therefore, we
infer that quantitative muscle ultrasonography has the po-
tential to predict the occurrence (and the evolution) of a
myopathic process. Moreover, we strongly feel that com-
binations of different texture features may be required for
prediction, diagnosis, monitoring and prognosis of the
myopathic disorders.
This work has the following two main limitations.
From the practical point of view, this feature-based char-
acterization is currently still not available in commercial
scanners. However, because the computation of the fea-
tures is not very demanding, it is likely that applications
for the characterization of images based on higher-order
features will be embedded in high-end scanners in a close
future. A second limitation is relative to the actual capa-
bility of detecting and grading pathologic conditions by
using this multi-features approach. This study is meth-
odologic and aimed at demonstrating the higher
sensitivity of higher-order features compared with the
first-order ones in characterizing the different muscles.
We are currently performing human investigations
focused on the quantification of sarcopenia and myostea-
tosis based on this technique.
CONCLUSIONS
We have proposed a texture-based technique to
differentiate gender and muscle types. In this work, we
found that the combination of first-order and higher-
order texture descriptors (Haralick, Galloway and LBP)
help to discriminate gender and muscle type. The most
important features included entropy, because the local
or global homogeneity of the image is clinically corre-
lated to the physiologic muscle status. The presented
technique is objective and non-invasive, and the prelimi-
nary results indicate the possibility of using this tech-
nique to differentiate gender and muscle types. Hence,
in future multi-texture analysis can be used to study the
muscle damage and myopathic disorders.
 Muscle ultrasonography texture analysis d F. MOLINARI et al.
AcknowledgmentsThis study was supported by the bank foundation
Fondazione CARIPLO of Milano, Italy (Project: Steroid myopathy:
Molecular, Histopathological, and Electrophysiological Characteriza-
tion) and by a grant (ex 60%) from the University of Turin.
REFERENCES
Acharya UR, Faust O, Alvin AP, Sree SV, Molinari F, Saba L,
Nicolaides A, Suri JS. Symptomatic vs. asymptomatic plaque clas-
sification in carotid ultrasound. J Med Syst 2012a;36:18611871.
Acharya UR, Faust O, Molinari F, Sree SV, Junnarkar SP, Sudarshan V.
Ultrasound-based tissue characterization and classification of fatty
liver disease: A screening and diagnostic paradigm. Knowl-Based
Syst 2015;75:6677.
Acharya UR, Faust O, Sree SV, Molinari F, Suri JS. ThyroScreen sys-
tem: High resolution ultrasound thyroid image characterization
into benign and malignant classes using novel combination of
texture and discrete wavelet transform. Comput Methods Programs
Biomed 2012b;107:233241.
Acharya UR, Mookiah MR, Vinitha Sree S, Yanti R, Martis RJ, Saba L,
Molinari F, Guerriero S, Suri JS. Evolutionary algorithm-based clas-
sifier parameter tuning for automatic ovarian cancer tissue character-
ization and classification. Ultraschall Med 2012c;35:237245.
Acharya UR, Sree SV, Krishnan MM, Molinari F, Saba L, Ho SY,
Ahuja AT, Ho SC, Nicolaides A, Suri JS. Atherosclerotic risk strat-
ification strategy for carotid arteries using texture-based features.
Ultrasound Med Biol 2012d;38:899915.
Acharya UR, Sree SV, Krishnan MM, Molinari F, Zieleznik W,
Bardales RH, Witkowska A, Suri JS. Computer-aided diagnostic
system for detection of Hashimoto thyroiditis on ultrasound images
from a Polish population. J Ultrasound Med 2014a;33:245253.
Acharya UR, Sree SV, Krishnan MM, Saba L, Molinari F, Guerriero S,
Suri JS. Ovarian tumor characterization using 3 D ultrasound. Tech-
nol Cancer Res Treat 2012e;11:543552.
Acharya UR, Sree SV, Saba L, Molinari F, Guerriero S, Suri JS. Ovarian
tumor characterization and classification using ultrasounda new
online paradigm. J Digital Imaging 2013;26:544553.
Acharya UR, Swapna G, Sree SV, Molinari F, Gupta S, Bardales RH,
Witkowska A, Suri JS. A review on ultrasound-based thyroid cancer
tissue characterization and automated classification. Technol Cancer
Res Treat 2014b;13:289301.
Acharya UR, Vinitha Sree S, Krishnan MM, Molinari F, Garberoglio R,
Suri JS. Non-invasive automated 3 D thyroid lesion classification in
ultrasound: A class of ThyroScan systems. Ultrasonics 2012f;52:
508520.
Arts IM, Pillen S, Schelhaas HJ, Overeem S, Zwarts MJ. Normal values
for quantitative muscle ultrasonography in adults. Muscle Nerve
2010;41:3241.
Caresio C, Molinari F, Emanuel G, Minetto MA. Muscle echo intensity:
Reliability and conditioning factors. Clin Physiol Funct Imaging
2014 [Epub ahead of print].
Chow RS, Medri MK, Martin DC, Leekam RN, Agur AM, McKee NH.
Sonographic studies of human soleus and gastrocnemius muscle ar-
chitecture: Gender variability. Eur J Appl Physiol 2000;82:236244.
Costanza MC, Afifi AA. Comparison of stopping rules in forward step-
wise discriminant analysis. J Am Stat Assoc 1979;74:777785.
Elias LJ, Bryden MP, Bulman-Fleming MB. Footedness is a better pre-
dictor than is handedness of emotional lateralization. Neuropsycho-
logia 1998;36:3743.
Fujikake T, Hart R, Nosaka K. Changes in B-mode ultrasound echo in-
tensity following injection of bupivacaine hydrochloride to rat hind
limb muscles in relation to histologic changes. Ultrasound Med Biol
2009;35:687696.
Galloway MM. Texture analysis using gray level run lengths. Comp
Graphics Image Process 1975;4:172179.
Gao S, Peng Y, Guo H, Liu W, Gao T, Xu Y, Tang X. Texture analysis
and classification of ultrasound liver images. Biomed Mater Eng
2014;24:12091216.
2531
Gdynia HJ, Muller HP, Ludolph AC, Koninger H, Huber R. Quantitative
muscle ultrasound in neuromuscular disorders using the parameters
intensity, entropy, and fractal dimension. Eur J Neurol 2009;
16:11511158.
Haralick RM, Shanmugam K, Dinstein IH. Textural features for image
classification. IEEE Trans 1973;SMC-3:610621.
Kim N, Amin V, Wilson D, Rouse G, Udpa S. Ultrasound image texture
analysis for characterizing intramuscular fat content of live beef cat-
tle. Ultrason Imaging 1998;20:191205.
Maurits NM, Beenakker EA, van Schaik DE, Fock JM, van der
Hoeven JH. Muscle ultrasound in children: Normal values and appli-
cation to neuromuscular disorders. Ultrasound Med Biol 2004;30:
10171027.
Maurits NM, Bollen AE, Windhausen A, De Jager AEJ, Van Der
Hoeven JH. Muscle ultrasound analysis: Normal values and differ-
entiation between myopathies and neuropathies. Ultrasound Med
Biol 2003;29:215225.
Narici MV, Binzoni T, Hiltbrand E, Fasel J, Terrier F, Cerretelli P. In vivo
human gastrocnemius architecture with changing joint angle at rest
and during graded isometric contraction. J Physiol 1996;496(Pt 1):
287297.
Niu L, Qian M, Yang W, Meng L, Xiao Y, Wong KK, Abbott D, Liu X,
Zheng H. Surface roughness detection of arteries via texture analysis
of ultrasound images for early diagnosis of atherosclerosis. PloS
One 2013;8:e76880.
Ojala T, Pietikainen M, Harwood D. A comparative study of texture
measures with classification based on featured distributions. Pattern
Recognition 1996;29:5159.
Ojala T, Pietikainen M, Maenpaa T. Multiresolution gray-scale and rota-
tion invariant texture classification with local binary patterns. IEEE
Trans 2002;24:971987.
Pillen S, Arts IM, Zwarts MJ. Muscle ultrasound in neuromuscular dis-
orders. Muscle Nerve 2008;37:679693.
Pillen S, Tak RO, Zwarts MJ, Lammens MM, Verrijp KN, Arts IM, van
der Laak JA, Hoogerbrugge PM, van Engelen BG, Verrips A. Skel-
etal muscle ultrasound: Correlation between fibrous tissue and echo
intensity. Ultrasound Med Biol 2009a;35:443446.
Pillen S, van Alfen N. Skeletal muscle ultrasound. Neurol Res 2011;33:
10161024.
Pillen S, van Dijk JP, Weijers G, Raijmann W, de Korte CL, Zwarts MJ.
Quantitative gray-scale analysis in skeletal muscle ultrasound: A
comparison study of two ultrasound devices. Muscle Nerve 2009b;
39:781786.
Radaelli R, Bottaro M, Wilhelm EN, Wagner DR, Pinto RS. Time course
of strength and echo intensity recovery after resistance exercise in
women. J Strength Cond Res 2012;26:25772584.
Reeves ND, Maganaris CN, Narici MV. Ultrasonographic assessment of
human skeletal muscle size. Eur J Appl Physiol 2004;91:116118.
Singh A, Singh B. Texture features extraction in mammograms using
non-shannon entropies. In: Ao S-I, Rieger B, Amouzegar MA,
(eds). Machine learning and systems engineering. Dordrecht, The
Netherlands: Springer; 2010. p. 341351.
Turo D, Otto P, Shah JP, Heimur J, Gebreab T, Zaazhoa M, Armstrong K,
Gerber LH, Sikdar S. Ultrasonic characterization of the upper trape-
zius muscle in patients with chronic neck pain. Ultrason Imaging
2013;35:173187.
Walker FO, Cartwright MS, Wiesler ER, Caress J. Ultrasound of nerve
and muscle. Clin Neurophysiol 2004;115:495507.
Zaidman CM, Connolly AM, Malkus EC, Florence JM, Pestronk A.
Quantitative ultrasound using backscatter analysis in Duchenne
and Becker muscular dystrophy. Neuromuscul Disord 2010;20:
805809.
Zaidman CM, Holland MR, Anderson CC, Pestronk A. Calibrated quan-
titative ultrasound imaging of skeletal muscle using backscatter
analysis. Muscle Nerve 2008;38:893898.
Zaidman CM, Holland MR, Hughes MS. Quantitative ultrasound of
skeletal muscle: reliable measurements of calibrated muscle back-
scatter from different ultrasound systems. Ultrasound Med Biol
2012;38:16181625.
2532 Ultrasound in Medicine and Biology Volume 41, Number 9, 2015

APPENDIX LOCAL BINARY PATTERN FEATURES

The LBP is computed by using the following method:

HARALICK FEATURES 1. A circular neighborhood of radius R pixels is considered around a

pixel. The pixel is the center of the circular neighborhood and it
The Haralick features are based on the gray level co-occurrence has intensity equal to Ic.
matrix (GLCM). Let the image be represented by a M 3 N gray-scale 2. P points are chosen on the circumference of the circle with radius R
matrix I(x,y), where each element of the matrix indicates the intensity such that they are all equidistant. Let Ip (p 5 1 . P) be the intensities
of a single pixel in the image. The co-occurrence matrix C(i,j|Dx, Dy) of the P points on the circumference.
is the second-order probability function estimation. This matrix de- 3. These P pixels are converted into a circular bit-stream of 0s and 1s
notes the rate of occurrence of a pixel pair with gray levels i and according to whether the gray value of the pixel is less than or greater
j, given the distances between the pixels are Dx and Dy in the x than Ic.
and y directions, respectively. The co-occurrence matrix C (i, j|Dx,
Dy) is defined as We considered P equal to 24 pixels and R equal to 3 pixels, in or-
der to consider a relatively large neighborhood. The created neighbor-
Ci; jjDx; Dy 5 jfp; q; p1Dx; q1Dy hood is then assigned with a uniformity measurement U that counts
the number of bit transitions (from 0 to 1 and vice versa) in the circular
: Ip; q 5 i; Ip1Dx; q1Dy 5 jgj (A1)
domain (with reference to Fig. 3, the U value would be equal to 4). We
assigned to the LBP code only the patterns with U # 2, so that:
where p; qp1Dx; q1DyMxN, d 5 Dx; Dy, and j$j denotes 8
the cardinality of a set. The probability that a gray level pixel <P P  
s I 2I Ux#2
i is at a distance (Dx, Dy) away from the gray level pixel j is LBPx 5 p 5 1 p c (A3)
given by :
P11 otherwise
Ci; j
Pi; j 5 P (A2) where s(x) is the step function that equals 1 if x $ 0 and is null for x , 0.
Ci; j
Lets fi be the relative frequency of the histogram derived from the LBP
values of all the image pixels. The LBP energy is defined as:
An element of the GLCM matrix (i, j, d, q) is defined as the joint X
probability of the gray levels i and j separated by distance d and along LBPenergy 5 fi2 (A4)
angular direction q. To reduce the computation burden, we have consid- i
ered q as 0 , 45 , 90 and 135 , and d is defined as the Manhattan or city
block distance (i.e. the number of pixels that must be crossed) based on The LBP entropy can be defined as:
this GLCM (Haralick et al. 1973). These second-order features are X
mathematically defined in Table 2. LBPentropy 5 2 fi log2 fi (A5)
i